audia 0.4.4

An agentic Python package that converts ideas and documents into audio – PDF papers, reports, and regulations turned into podcast-style audio files.

audia — turn your ideas into audio

audia is an agentic Python package that converts PDFs — academic papers, reports, regulations — into podcast-style audio files. It uses an LLM to rewrite content into natural spoken language (math in plain English, tables as sentences, no citations) before passing it to a TTS engine, so the result actually sounds good when read aloud.

The audia CLI

Features

• LLM-curated text — mandatory LLM pass rewrites math notation, condenses tables and acknowledgements, removes citation artefacts, and ensures smooth spoken flow
• Chunk-level stitching — long documents are split at paragraph boundaries; each chunk receives the tail of the previous curated output as transition context
• ArXiv research — search papers by query and convert them to audio in one command
• Voice input (STT) — record a spoken query to trigger an ArXiv search
• Multiple TTS backends — edge-tts (default, free), kokoro (local), or OpenAI TTS
• Multiple LLM backends — OpenAI (gpt-4o-mini default) or Anthropic
• CLI — audia convert, research, listen, serve, info
• Web UI — FastAPI backend + SPA frontend
• Local storage — SQLite database for papers and audio files via SQLAlchemy
• Debug output — every run saves raw, preprocessed, and curated text to ~/.audia/debug/<run_id>/

Tech Stack

Backend

• Python 3.10+ — package language
• FastAPI — backend for the web UI
• LangGraph — agentic pipeline orchestration (PDF → preprocess → LLM curate → TTS)
• LangChain — LLM abstraction
  • Current support for LLMs from Anthropic, Google, and OpenAI
• edge-tts — default TTS backend, no API key required
• faster-whisper — STT for voice input
• PyMuPDF — PDF text extraction
• SQLite — local database for papers and audio files

Frontend

• React — interactive frontend
• Vite — fast dev server and production bundler
• Tailwind CSS — utility-first styling
• TypeScript — type-safe component and API code

CLI

• Typer + Rich — CLI with coloured progress output

Packaging

• PyPI — distributed as an installable Python package

Installation

pip install audia

For CLI usage, pipx is recommended — it installs audia in an isolated environment while exposing the command globally:

pipx install audia

Optional extras:

Extra	Installs
kokoro	local Kokoro TTS

pip install "audia[kokoro]"

Configuration

Copy .env.example to .env in your working directory and set your API key:

cp .env.example .env

Minimum required settings:

AUDIA_LLM_PROVIDER=openai           # or anthropic
AUDIA_OPENAI_API_KEY=sk-...

All settings use the AUDIA_ prefix. Run audia info to see the active configuration.

Quick Start

Show active configuration:

audia info

Convert a local PDF:

audia convert paper.pdf

Convert multiple PDFs to a specific output folder:

audia convert paper1.pdf paper2.pdf --output ~/audiobooks

Search ArXiv and convert the top results:

audia research "retrieval augmented generation" --max-results 3 --convert

Start the web UI:

audia serve
# → http://localhost:8000

The audia UI

Pipeline

The pipeline can be entered in three ways:

Entry point	Command
Voice input	audia listen — record speech, LLM distils a search query, confirm, then runs the full pipeline
Text query	audia research "retrieval augmented generation" — search ArXiv by text, select papers, run pipeline
Local PDF	audia convert paper.pdf — skip Steps 0, go straight to extraction

When starting from voice or text, the full five-step LangGraph pipeline runs. For local PDFs, Steps 1–4 run directly:

 [voice input]          [text query]
      │                      │
      ▼                      │
  Microphone                 │
  (faster-whisper STT)       │
      │                      │
      ▼                      │
  LLM query distillation     │        ← extracts concise ArXiv search terms
      │                      │           from natural speech
      ▼                      │
  Confirm / re-record?       │
      │  yes                 │
      ▼                      ▼
Step 0 — ArXiv search    (or use local PDF)
 │        arxiv API: fetch metadata, download PDF
 │
 ▼
Step 1 — PDF extraction       PyMuPDF: text + metadata per page
 │
 ▼
Step 2 — Heuristic pre-pass   Regex: strip citations, LaTeX commands, figure captions
 │
 ▼
Step 3 — LLM curation         Chunked LLM pass: math → English, tables → sentences,
 │                             smooth spoken transitions between chunks
 ▼
Step 4 — TTS synthesis        edge-tts (or kokoro / OpenAI): split into ~3800-char
                               chunks, synthesise, concatenate → .mp3

Output files for a run on 2025_Xu+.pdf:

~/.audia/audio/2025_Xu+_20260329_084445.mp3
~/.audia/debug/2025_Xu+_20260329_084445/
    1_raw.txt            ← PyMuPDF output
    2_preprocessed.txt   ← after heuristic pass
    3_curated.txt        ← after LLM curation

Contributing

1. Fork the repository
2. Create a feature branch (git checkout -b feature/my-change)
3. Make your changes
4. Run the test suite: pytest --cov=src --cov-report=term-missing
5. Submit a pull request

License

MIT — see LICENSE for details.